Expose GitHub CLI through the agent github tool

[nishu-builder]

Expose github(repository, args, stdin?) through std/github, using lazy worker evaluation and the existing callback path. Share repository validation with the worker and label uncertain results prominently so they are not treated as safe-to-retry command failures.

Extend the branch design with GitHub invocation recovery, PR creation/updates and stack workflows using existing PR URLs. Retain the full stored request across recovery, including its worker and credential identity.

The fixture covers registration, secret grants, lazy evaluation, callback, version/help and replay. Live PR writes and stack linking remain follow-up validation.

Validation on EC2: Nix build passed. The salted suite passed 46/47 checks in 395s; its Clippy findings were fixed, and Clippy passed on all affected crates with --all-targets -- -D warnings. Formatting passed after those lint-only fixes. HTTPS import/push, shallow-locator/offline-cache, and focused unit checks also passed. The functional suite was not repeated for the lint-only changes.

[nishu-builder]

Layer 7/8. The bind-as-data / evaluate-when-called trick via std/llm-step/github/DEPS is a neat way to keep the secret-marked worker out of the agent's own reader identity, and the comment explains why. The invocation id derived from (conversation, request, round, call) gives exactly the "new call, new id; resumed call, same id" the design wants, and resume goes through the stored task so a rebuilt image does not re-key a half-done call. The stub-driven tests/github extension checks the registry, the callback and the source_tree parameter on publish_source, which closes the "registry silently drops a tool" hole from CLAUDE.md.

Two minor inline comments (duplicated repository validation, and making uncertain visually distinct from a plain non-zero exit in the result block). Nothing blocking.

---

Generated by Claude Code

[nishu-builder]

Closing with the rest of stack #261. #300 replaces this: the agent's github tool runs inline in llm-step against the REST API, with no std/github worker or gh. The implementation follows in a PR stacked on #300.

---

Generated by Claude Code